unknown

Liquid AI has released LFM2.5-1.2B-Thinking, a compact 1.2 billion parameter reasoning model that runs entirely on-device using less than 1 GB of memory. The model features a 32,768 token context window and generates explicit thinking traces before producing final answers, making it suitable for agents, tool use, mathematics, and RAG workflows. It achieves strong benchmark results including 87.96 on MATH 500 and 85.60 on GSM8K, while a multi-stage training pipeline with supervised reasoning traces, preference alignment, and RLVR dramatically reduces doom looping from 15.74% to 0.36%.

Liquid AI Releases LFM2.5-1.2B-Thinking: a 1.2B Parameter Reasoning Model That Fits Under 1 GB On-Device Liquid AI releases LFM2.5-1.2B-Thinking, a 1.2 billion parameter reasoning model that runs…

We are a community of AI/ ML/Generative AI enthusiasts/researchers/journalists/writers who share interesting news and articles about the applications of AI. 